0041: Bound Visual Questions Before Inventory
Status: Accepted; routing folded into ADR 0048
Date: 2026-07-22
Context
ADR 0040 prevented generic human or AI review from justifying detached capture infrastructure. A 0.5.7 candidate behavioral probe obeyed that artifact rule, but still performed a broad source audit for a request to inspect every visual object. It converted the audit into proposed subject taxonomies, scenario families, coverage profiles, and expansion of an existing project Almanac.
The final recommendation stayed native, but the workflow was still driven by inventory completion rather than a visual decision. The skill's sequence made that likely: it asked the agent to orient in the project before requiring one bounded question.
Decision
- Dedicated visual inspection begins by naming one subject, state or beat, material context, and failure condition.
- If a request is unbounded, stop before repository-wide visual inventory and ask for one decision-bearing question.
- Inspect only the production paths material to that question.
- Do not translate "every visual object" into a coverage taxonomy, scenario catalog, fixed matrix, or Almanac expansion plan.
- An existing gallery, codex, or Almanac is a possible project-owned route, not an obligation to cover every asset or presentation surface.
- Broad implementation work with a concrete product outcome remains ordinary
stage-deliver-changework; this decision constrains dedicated visual diagnosis, not legitimate asset audits or feature backlogs requested by the owner.
Consequences
- Agents may ask for scope before reading large parts of a repository.
- Visual investigation becomes cheaper and less likely to manufacture work.
- Existing project viewers remain useful without becoming STAGE conformance surfaces.
- Comprehensive audits remain possible when the owner explicitly names their product or release decision and accepts their cost.
Evidence Boundary
This decision comes from one deliberately adversarial probe against the same-owner Lanternworks dogfood project. It demonstrates a routing failure in the installed skill, not that broad art audits are universally wasteful.